| 1234567891011121314151617181920212223242526272829 | from torch import nnclass LRASPP(nn.Module):    def __init__(self, in_channels, out_channels):        super().__init__()        self.aspp1 = nn.Sequential(            nn.Conv2d(in_channels, out_channels, 1, bias=False),            nn.BatchNorm2d(out_channels),            nn.ReLU(True)        )        self.aspp2 = nn.Sequential(            nn.AdaptiveAvgPool2d(1),            nn.Conv2d(in_channels, out_channels, 1, bias=False),            nn.Sigmoid()        )            def forward_single_frame(self, x):        return self.aspp1(x) * self.aspp2(x)        def forward_time_series(self, x):        B, T = x.shape[:2]        x = self.forward_single_frame(x.flatten(0, 1)).unflatten(0, (B, T))        return x        def forward(self, x):        if x.ndim == 5:            return self.forward_time_series(x)        else:            return self.forward_single_frame(x)
 |